2 * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
3 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
4 * Copyright (c) 2013-2014 Mellanox Technologies. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/kernel.h>
35 #include <linux/module.h>
36 #include <linux/slab.h>
37 #include <linux/delay.h>
39 #include "iscsi_iser.h"
41 #define ISCSI_ISER_MAX_CONN 8
42 #define ISER_MAX_RX_LEN (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
43 #define ISER_MAX_TX_LEN (ISER_QP_MAX_REQ_DTOS * ISCSI_ISER_MAX_CONN)
44 #define ISER_MAX_CQ_LEN (ISER_MAX_RX_LEN + ISER_MAX_TX_LEN + \
47 static int iser_cq_poll_limit = 512;
49 static void iser_cq_tasklet_fn(unsigned long data);
50 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
52 static void iser_cq_event_callback(struct ib_event *cause, void *context)
54 iser_err("got cq event %d \n", cause->event);
57 static void iser_qp_event_callback(struct ib_event *cause, void *context)
59 iser_err("got qp event %d\n",cause->event);
62 static void iser_event_handler(struct ib_event_handler *handler,
63 struct ib_event *event)
65 iser_err("async event %d on device %s port %d\n", event->event,
66 event->device->name, event->element.port_num);
70 * iser_create_device_ib_res - creates Protection Domain (PD), Completion
71 * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
74 * returns 0 on success, -1 on failure
76 static int iser_create_device_ib_res(struct iser_device *device)
78 struct ib_device_attr *dev_attr = &device->dev_attr;
81 ret = ib_query_device(device->ib_device, dev_attr);
83 pr_warn("Query device failed for %s\n", device->ib_device->name);
87 /* Assign function handles - based on FMR support */
88 if (device->ib_device->alloc_fmr && device->ib_device->dealloc_fmr &&
89 device->ib_device->map_phys_fmr && device->ib_device->unmap_fmr) {
90 iser_info("FMR supported, using FMR for registration\n");
91 device->iser_alloc_rdma_reg_res = iser_create_fmr_pool;
92 device->iser_free_rdma_reg_res = iser_free_fmr_pool;
93 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fmr;
94 device->iser_unreg_rdma_mem = iser_unreg_mem_fmr;
96 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
97 iser_info("FastReg supported, using FastReg for registration\n");
98 device->iser_alloc_rdma_reg_res = iser_create_fastreg_pool;
99 device->iser_free_rdma_reg_res = iser_free_fastreg_pool;
100 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fastreg;
101 device->iser_unreg_rdma_mem = iser_unreg_mem_fastreg;
103 iser_err("IB device does not support FMRs nor FastRegs, can't register memory\n");
107 device->comps_used = min_t(int, num_online_cpus(),
108 device->ib_device->num_comp_vectors);
110 device->comps = kcalloc(device->comps_used, sizeof(*device->comps),
115 max_cqe = min(ISER_MAX_CQ_LEN, dev_attr->max_cqe);
117 iser_info("using %d CQs, device %s supports %d vectors max_cqe %d\n",
118 device->comps_used, device->ib_device->name,
119 device->ib_device->num_comp_vectors, max_cqe);
121 device->pd = ib_alloc_pd(device->ib_device);
122 if (IS_ERR(device->pd))
125 for (i = 0; i < device->comps_used; i++) {
126 struct iser_comp *comp = &device->comps[i];
128 comp->device = device;
129 comp->cq = ib_create_cq(device->ib_device,
131 iser_cq_event_callback,
134 if (IS_ERR(comp->cq)) {
139 if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
142 tasklet_init(&comp->tasklet, iser_cq_tasklet_fn,
143 (unsigned long)comp);
146 device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
147 IB_ACCESS_REMOTE_WRITE |
148 IB_ACCESS_REMOTE_READ);
149 if (IS_ERR(device->mr))
152 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
154 if (ib_register_event_handler(&device->event_handler))
160 ib_dereg_mr(device->mr);
162 for (i = 0; i < device->comps_used; i++)
163 tasklet_kill(&device->comps[i].tasklet);
165 for (i = 0; i < device->comps_used; i++) {
166 struct iser_comp *comp = &device->comps[i];
169 ib_destroy_cq(comp->cq);
171 ib_dealloc_pd(device->pd);
173 kfree(device->comps);
175 iser_err("failed to allocate an IB resource\n");
180 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
181 * CQ and PD created with the device associated with the adapator.
183 static void iser_free_device_ib_res(struct iser_device *device)
186 BUG_ON(device->mr == NULL);
188 for (i = 0; i < device->comps_used; i++) {
189 struct iser_comp *comp = &device->comps[i];
191 tasklet_kill(&comp->tasklet);
192 ib_destroy_cq(comp->cq);
196 (void)ib_unregister_event_handler(&device->event_handler);
197 (void)ib_dereg_mr(device->mr);
198 (void)ib_dealloc_pd(device->pd);
200 kfree(device->comps);
201 device->comps = NULL;
208 * iser_create_fmr_pool - Creates FMR pool and page_vector
210 * returns 0 on success, or errno code on failure
212 int iser_create_fmr_pool(struct ib_conn *ib_conn, unsigned cmds_max)
214 struct iser_device *device = ib_conn->device;
215 struct ib_fmr_pool_param params;
218 ib_conn->fmr.page_vec = kmalloc(sizeof(*ib_conn->fmr.page_vec) +
219 (sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
221 if (!ib_conn->fmr.page_vec)
224 ib_conn->fmr.page_vec->pages = (u64 *)(ib_conn->fmr.page_vec + 1);
226 params.page_shift = SHIFT_4K;
227 /* when the first/last SG element are not start/end *
228 * page aligned, the map whould be of N+1 pages */
229 params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
230 /* make the pool size twice the max number of SCSI commands *
231 * the ML is expected to queue, watermark for unmap at 50% */
232 params.pool_size = cmds_max * 2;
233 params.dirty_watermark = cmds_max;
235 params.flush_function = NULL;
236 params.access = (IB_ACCESS_LOCAL_WRITE |
237 IB_ACCESS_REMOTE_WRITE |
238 IB_ACCESS_REMOTE_READ);
240 ib_conn->fmr.pool = ib_create_fmr_pool(device->pd, ¶ms);
241 if (!IS_ERR(ib_conn->fmr.pool))
244 /* no FMR => no need for page_vec */
245 kfree(ib_conn->fmr.page_vec);
246 ib_conn->fmr.page_vec = NULL;
248 ret = PTR_ERR(ib_conn->fmr.pool);
249 ib_conn->fmr.pool = NULL;
250 if (ret != -ENOSYS) {
251 iser_err("FMR allocation failed, err %d\n", ret);
254 iser_warn("FMRs are not supported, using unaligned mode\n");
260 * iser_free_fmr_pool - releases the FMR pool and page vec
262 void iser_free_fmr_pool(struct ib_conn *ib_conn)
264 iser_info("freeing conn %p fmr pool %p\n",
265 ib_conn, ib_conn->fmr.pool);
267 if (ib_conn->fmr.pool != NULL)
268 ib_destroy_fmr_pool(ib_conn->fmr.pool);
270 ib_conn->fmr.pool = NULL;
272 kfree(ib_conn->fmr.page_vec);
273 ib_conn->fmr.page_vec = NULL;
277 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
278 bool pi_enable, struct fast_reg_descriptor *desc)
282 desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
283 ISCSI_ISER_SG_TABLESIZE + 1);
284 if (IS_ERR(desc->data_frpl)) {
285 ret = PTR_ERR(desc->data_frpl);
286 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
288 return PTR_ERR(desc->data_frpl);
291 desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
292 if (IS_ERR(desc->data_mr)) {
293 ret = PTR_ERR(desc->data_mr);
294 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
295 goto fast_reg_mr_failure;
297 desc->reg_indicators |= ISER_DATA_KEY_VALID;
300 struct ib_mr_init_attr mr_init_attr = {0};
301 struct iser_pi_context *pi_ctx = NULL;
303 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
305 iser_err("Failed to allocate pi context\n");
307 goto pi_ctx_alloc_failure;
309 pi_ctx = desc->pi_ctx;
311 pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
312 ISCSI_ISER_SG_TABLESIZE);
313 if (IS_ERR(pi_ctx->prot_frpl)) {
314 ret = PTR_ERR(pi_ctx->prot_frpl);
315 iser_err("Failed to allocate prot frpl ret=%d\n",
317 goto prot_frpl_failure;
320 pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
321 ISCSI_ISER_SG_TABLESIZE + 1);
322 if (IS_ERR(pi_ctx->prot_mr)) {
323 ret = PTR_ERR(pi_ctx->prot_mr);
324 iser_err("Failed to allocate prot frmr ret=%d\n",
326 goto prot_mr_failure;
328 desc->reg_indicators |= ISER_PROT_KEY_VALID;
330 mr_init_attr.max_reg_descriptors = 2;
331 mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
332 pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
333 if (IS_ERR(pi_ctx->sig_mr)) {
334 ret = PTR_ERR(pi_ctx->sig_mr);
335 iser_err("Failed to allocate signature enabled mr err=%d\n",
339 desc->reg_indicators |= ISER_SIG_KEY_VALID;
341 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
343 iser_dbg("Create fr_desc %p page_list %p\n",
344 desc, desc->data_frpl->page_list);
348 ib_dereg_mr(desc->pi_ctx->prot_mr);
350 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
353 pi_ctx_alloc_failure:
354 ib_dereg_mr(desc->data_mr);
356 ib_free_fast_reg_page_list(desc->data_frpl);
362 * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
363 * for fast registration work requests.
364 * returns 0 on success, or errno code on failure
366 int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max)
368 struct iser_device *device = ib_conn->device;
369 struct fast_reg_descriptor *desc;
372 INIT_LIST_HEAD(&ib_conn->fastreg.pool);
373 ib_conn->fastreg.pool_size = 0;
374 for (i = 0; i < cmds_max; i++) {
375 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
377 iser_err("Failed to allocate a new fast_reg descriptor\n");
382 ret = iser_create_fastreg_desc(device->ib_device, device->pd,
383 ib_conn->pi_support, desc);
385 iser_err("Failed to create fastreg descriptor err=%d\n",
391 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
392 ib_conn->fastreg.pool_size++;
398 iser_free_fastreg_pool(ib_conn);
403 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
405 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
407 struct fast_reg_descriptor *desc, *tmp;
410 if (list_empty(&ib_conn->fastreg.pool))
413 iser_info("freeing conn %p fr pool\n", ib_conn);
415 list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
416 list_del(&desc->list);
417 ib_free_fast_reg_page_list(desc->data_frpl);
418 ib_dereg_mr(desc->data_mr);
420 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
421 ib_dereg_mr(desc->pi_ctx->prot_mr);
422 ib_destroy_mr(desc->pi_ctx->sig_mr);
429 if (i < ib_conn->fastreg.pool_size)
430 iser_warn("pool still has %d regions registered\n",
431 ib_conn->fastreg.pool_size - i);
435 * iser_create_ib_conn_res - Queue-Pair (QP)
437 * returns 0 on success, -1 on failure
439 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
441 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
443 struct iser_device *device;
444 struct ib_device_attr *dev_attr;
445 struct ib_qp_init_attr init_attr;
447 int index, min_index = 0;
449 BUG_ON(ib_conn->device == NULL);
451 device = ib_conn->device;
452 dev_attr = &device->dev_attr;
454 memset(&init_attr, 0, sizeof init_attr);
456 mutex_lock(&ig.connlist_mutex);
457 /* select the CQ with the minimal number of usages */
458 for (index = 0; index < device->comps_used; index++) {
459 if (device->comps[index].active_qps <
460 device->comps[min_index].active_qps)
463 ib_conn->comp = &device->comps[min_index];
464 ib_conn->comp->active_qps++;
465 mutex_unlock(&ig.connlist_mutex);
466 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
468 init_attr.event_handler = iser_qp_event_callback;
469 init_attr.qp_context = (void *)ib_conn;
470 init_attr.send_cq = ib_conn->comp->cq;
471 init_attr.recv_cq = ib_conn->comp->cq;
472 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
473 init_attr.cap.max_send_sge = 2;
474 init_attr.cap.max_recv_sge = 1;
475 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
476 init_attr.qp_type = IB_QPT_RC;
477 if (ib_conn->pi_support) {
478 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
479 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
480 iser_conn->max_cmds =
481 ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
483 if (dev_attr->max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
484 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS + 1;
485 iser_conn->max_cmds =
486 ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
488 init_attr.cap.max_send_wr = dev_attr->max_qp_wr;
489 iser_conn->max_cmds =
490 ISER_GET_MAX_XMIT_CMDS(dev_attr->max_qp_wr);
491 iser_dbg("device %s supports max_send_wr %d\n",
492 device->ib_device->name, dev_attr->max_qp_wr);
496 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
500 ib_conn->qp = ib_conn->cma_id->qp;
501 iser_info("setting conn %p cma_id %p qp %p\n",
502 ib_conn, ib_conn->cma_id,
503 ib_conn->cma_id->qp);
507 mutex_lock(&ig.connlist_mutex);
508 ib_conn->comp->active_qps--;
509 mutex_unlock(&ig.connlist_mutex);
510 iser_err("unable to alloc mem or create resource, err %d\n", ret);
516 * based on the resolved device node GUID see if there already allocated
517 * device for this device. If there's no such, create one.
520 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
522 struct iser_device *device;
524 mutex_lock(&ig.device_list_mutex);
526 list_for_each_entry(device, &ig.device_list, ig_list)
527 /* find if there's a match using the node GUID */
528 if (device->ib_device->node_guid == cma_id->device->node_guid)
531 device = kzalloc(sizeof *device, GFP_KERNEL);
535 /* assign this device to the device */
536 device->ib_device = cma_id->device;
537 /* init the device and link it into ig device list */
538 if (iser_create_device_ib_res(device)) {
543 list_add(&device->ig_list, &ig.device_list);
548 mutex_unlock(&ig.device_list_mutex);
552 /* if there's no demand for this device, release it */
553 static void iser_device_try_release(struct iser_device *device)
555 mutex_lock(&ig.device_list_mutex);
557 iser_info("device %p refcount %d\n", device, device->refcount);
558 if (!device->refcount) {
559 iser_free_device_ib_res(device);
560 list_del(&device->ig_list);
563 mutex_unlock(&ig.device_list_mutex);
567 * Called with state mutex held
569 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
570 enum iser_conn_state comp,
571 enum iser_conn_state exch)
575 ret = (iser_conn->state == comp);
577 iser_conn->state = exch;
582 void iser_release_work(struct work_struct *work)
584 struct iser_conn *iser_conn;
586 iser_conn = container_of(work, struct iser_conn, release_work);
588 /* Wait for conn_stop to complete */
589 wait_for_completion(&iser_conn->stop_completion);
590 /* Wait for IB resouces cleanup to complete */
591 wait_for_completion(&iser_conn->ib_completion);
593 mutex_lock(&iser_conn->state_mutex);
594 iser_conn->state = ISER_CONN_DOWN;
595 mutex_unlock(&iser_conn->state_mutex);
597 iser_conn_release(iser_conn);
601 * iser_free_ib_conn_res - release IB related resources
602 * @iser_conn: iser connection struct
603 * @destroy_device: indicator if we need to try to release
604 * the iser device (only iscsi shutdown and DEVICE_REMOVAL
607 * This routine is called with the iser state mutex held
608 * so the cm_id removal is out of here. It is Safe to
609 * be invoked multiple times.
611 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
614 struct ib_conn *ib_conn = &iser_conn->ib_conn;
615 struct iser_device *device = ib_conn->device;
617 iser_info("freeing conn %p cma_id %p qp %p\n",
618 iser_conn, ib_conn->cma_id, ib_conn->qp);
620 iser_free_rx_descriptors(iser_conn);
622 if (ib_conn->qp != NULL) {
623 ib_conn->comp->active_qps--;
624 rdma_destroy_qp(ib_conn->cma_id);
628 if (destroy_device && device != NULL) {
629 iser_device_try_release(device);
630 ib_conn->device = NULL;
635 * Frees all conn objects and deallocs conn descriptor
637 void iser_conn_release(struct iser_conn *iser_conn)
639 struct ib_conn *ib_conn = &iser_conn->ib_conn;
641 mutex_lock(&ig.connlist_mutex);
642 list_del(&iser_conn->conn_list);
643 mutex_unlock(&ig.connlist_mutex);
645 mutex_lock(&iser_conn->state_mutex);
646 if (iser_conn->state != ISER_CONN_DOWN) {
647 iser_warn("iser conn %p state %d, expected state down.\n",
648 iser_conn, iser_conn->state);
649 iser_conn->state = ISER_CONN_DOWN;
652 * In case we never got to bind stage, we still need to
653 * release IB resources (which is safe to call more than once).
655 iser_free_ib_conn_res(iser_conn, true);
656 mutex_unlock(&iser_conn->state_mutex);
658 if (ib_conn->cma_id != NULL) {
659 rdma_destroy_id(ib_conn->cma_id);
660 ib_conn->cma_id = NULL;
667 * triggers start of the disconnect procedures and wait for them to be done
668 * Called with state mutex held
670 int iser_conn_terminate(struct iser_conn *iser_conn)
672 struct ib_conn *ib_conn = &iser_conn->ib_conn;
673 struct ib_send_wr *bad_wr;
676 /* terminate the iser conn only if the conn state is UP */
677 if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
678 ISER_CONN_TERMINATING))
681 iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
683 /* suspend queuing of new iscsi commands */
684 if (iser_conn->iscsi_conn)
685 iscsi_suspend_queue(iser_conn->iscsi_conn);
688 * In case we didn't already clean up the cma_id (peer initiated
689 * a disconnection), we need to Cause the CMA to change the QP
692 if (ib_conn->cma_id) {
693 err = rdma_disconnect(ib_conn->cma_id);
695 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
698 /* post an indication that all flush errors were consumed */
699 err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
701 iser_err("conn %p failed to post beacon", ib_conn);
705 wait_for_completion(&ib_conn->flush_comp);
712 * Called with state mutex held
714 static void iser_connect_error(struct rdma_cm_id *cma_id)
716 struct iser_conn *iser_conn;
718 iser_conn = (struct iser_conn *)cma_id->context;
719 iser_conn->state = ISER_CONN_DOWN;
723 * Called with state mutex held
725 static void iser_addr_handler(struct rdma_cm_id *cma_id)
727 struct iser_device *device;
728 struct iser_conn *iser_conn;
729 struct ib_conn *ib_conn;
732 iser_conn = (struct iser_conn *)cma_id->context;
733 if (iser_conn->state != ISER_CONN_PENDING)
737 ib_conn = &iser_conn->ib_conn;
738 device = iser_device_find_by_ib_device(cma_id);
740 iser_err("device lookup/creation failed\n");
741 iser_connect_error(cma_id);
745 ib_conn->device = device;
747 /* connection T10-PI support */
748 if (iser_pi_enable) {
749 if (!(device->dev_attr.device_cap_flags &
750 IB_DEVICE_SIGNATURE_HANDOVER)) {
751 iser_warn("T10-PI requested but not supported on %s, "
752 "continue without T10-PI\n",
753 ib_conn->device->ib_device->name);
754 ib_conn->pi_support = false;
756 ib_conn->pi_support = true;
760 ret = rdma_resolve_route(cma_id, 1000);
762 iser_err("resolve route failed: %d\n", ret);
763 iser_connect_error(cma_id);
769 * Called with state mutex held
771 static void iser_route_handler(struct rdma_cm_id *cma_id)
773 struct rdma_conn_param conn_param;
775 struct iser_cm_hdr req_hdr;
776 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
777 struct ib_conn *ib_conn = &iser_conn->ib_conn;
778 struct iser_device *device = ib_conn->device;
780 if (iser_conn->state != ISER_CONN_PENDING)
784 ret = iser_create_ib_conn_res(ib_conn);
788 memset(&conn_param, 0, sizeof conn_param);
789 conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
790 conn_param.initiator_depth = 1;
791 conn_param.retry_count = 7;
792 conn_param.rnr_retry_count = 6;
794 memset(&req_hdr, 0, sizeof(req_hdr));
795 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
796 ISER_SEND_W_INV_NOT_SUPPORTED);
797 conn_param.private_data = (void *)&req_hdr;
798 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
800 ret = rdma_connect(cma_id, &conn_param);
802 iser_err("failure connecting: %d\n", ret);
808 iser_connect_error(cma_id);
811 static void iser_connected_handler(struct rdma_cm_id *cma_id)
813 struct iser_conn *iser_conn;
814 struct ib_qp_attr attr;
815 struct ib_qp_init_attr init_attr;
817 iser_conn = (struct iser_conn *)cma_id->context;
818 if (iser_conn->state != ISER_CONN_PENDING)
822 (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
823 iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
825 iser_conn->state = ISER_CONN_UP;
826 complete(&iser_conn->up_completion);
829 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
831 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
833 if (iser_conn_terminate(iser_conn)) {
834 if (iser_conn->iscsi_conn)
835 iscsi_conn_failure(iser_conn->iscsi_conn,
836 ISCSI_ERR_CONN_FAILED);
838 iser_err("iscsi_iser connection isn't bound\n");
842 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
845 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
848 * We are not guaranteed that we visited disconnected_handler
849 * by now, call it here to be safe that we handle CM drep
852 iser_disconnected_handler(cma_id);
853 iser_free_ib_conn_res(iser_conn, destroy_device);
854 complete(&iser_conn->ib_completion);
857 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
859 struct iser_conn *iser_conn;
862 iser_conn = (struct iser_conn *)cma_id->context;
863 iser_info("event %d status %d conn %p id %p\n",
864 event->event, event->status, cma_id->context, cma_id);
866 mutex_lock(&iser_conn->state_mutex);
867 switch (event->event) {
868 case RDMA_CM_EVENT_ADDR_RESOLVED:
869 iser_addr_handler(cma_id);
871 case RDMA_CM_EVENT_ROUTE_RESOLVED:
872 iser_route_handler(cma_id);
874 case RDMA_CM_EVENT_ESTABLISHED:
875 iser_connected_handler(cma_id);
877 case RDMA_CM_EVENT_ADDR_ERROR:
878 case RDMA_CM_EVENT_ROUTE_ERROR:
879 case RDMA_CM_EVENT_CONNECT_ERROR:
880 case RDMA_CM_EVENT_UNREACHABLE:
881 case RDMA_CM_EVENT_REJECTED:
882 iser_connect_error(cma_id);
884 case RDMA_CM_EVENT_DISCONNECTED:
885 case RDMA_CM_EVENT_ADDR_CHANGE:
886 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
887 iser_cleanup_handler(cma_id, false);
889 case RDMA_CM_EVENT_DEVICE_REMOVAL:
891 * we *must* destroy the device as we cannot rely
892 * on iscsid to be around to initiate error handling.
893 * also if we are not in state DOWN implicitly destroy
896 iser_cleanup_handler(cma_id, true);
897 if (iser_conn->state != ISER_CONN_DOWN) {
898 iser_conn->ib_conn.cma_id = NULL;
903 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
906 mutex_unlock(&iser_conn->state_mutex);
911 void iser_conn_init(struct iser_conn *iser_conn)
913 iser_conn->state = ISER_CONN_INIT;
914 iser_conn->ib_conn.post_recv_buf_count = 0;
915 init_completion(&iser_conn->ib_conn.flush_comp);
916 init_completion(&iser_conn->stop_completion);
917 init_completion(&iser_conn->ib_completion);
918 init_completion(&iser_conn->up_completion);
919 INIT_LIST_HEAD(&iser_conn->conn_list);
920 spin_lock_init(&iser_conn->ib_conn.lock);
921 mutex_init(&iser_conn->state_mutex);
925 * starts the process of connecting to the target
926 * sleeps until the connection is established or rejected
928 int iser_connect(struct iser_conn *iser_conn,
929 struct sockaddr *src_addr,
930 struct sockaddr *dst_addr,
933 struct ib_conn *ib_conn = &iser_conn->ib_conn;
936 mutex_lock(&iser_conn->state_mutex);
938 sprintf(iser_conn->name, "%pISp", dst_addr);
940 iser_info("connecting to: %s\n", iser_conn->name);
942 /* the device is known only --after-- address resolution */
943 ib_conn->device = NULL;
945 iser_conn->state = ISER_CONN_PENDING;
947 ib_conn->beacon.wr_id = ISER_BEACON_WRID;
948 ib_conn->beacon.opcode = IB_WR_SEND;
950 ib_conn->cma_id = rdma_create_id(iser_cma_handler,
952 RDMA_PS_TCP, IB_QPT_RC);
953 if (IS_ERR(ib_conn->cma_id)) {
954 err = PTR_ERR(ib_conn->cma_id);
955 iser_err("rdma_create_id failed: %d\n", err);
959 err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
961 iser_err("rdma_resolve_addr failed: %d\n", err);
966 wait_for_completion_interruptible(&iser_conn->up_completion);
968 if (iser_conn->state != ISER_CONN_UP) {
970 goto connect_failure;
973 mutex_unlock(&iser_conn->state_mutex);
975 mutex_lock(&ig.connlist_mutex);
976 list_add(&iser_conn->conn_list, &ig.connlist);
977 mutex_unlock(&ig.connlist_mutex);
981 ib_conn->cma_id = NULL;
983 iser_conn->state = ISER_CONN_DOWN;
985 mutex_unlock(&iser_conn->state_mutex);
986 iser_conn_release(iser_conn);
991 * iser_reg_page_vec - Register physical memory
993 * returns: 0 on success, errno code on failure
995 int iser_reg_page_vec(struct ib_conn *ib_conn,
996 struct iser_page_vec *page_vec,
997 struct iser_mem_reg *mem_reg)
999 struct ib_pool_fmr *mem;
1004 page_list = page_vec->pages;
1005 io_addr = page_list[0];
1007 mem = ib_fmr_pool_map_phys(ib_conn->fmr.pool,
1013 status = (int)PTR_ERR(mem);
1014 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
1018 mem_reg->lkey = mem->fmr->lkey;
1019 mem_reg->rkey = mem->fmr->rkey;
1020 mem_reg->len = page_vec->length * SIZE_4K;
1021 mem_reg->va = io_addr;
1022 mem_reg->mem_h = (void *)mem;
1024 mem_reg->va += page_vec->offset;
1025 mem_reg->len = page_vec->data_size;
1027 iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
1028 "entry[0]: (0x%08lx,%ld)] -> "
1029 "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
1030 page_vec, page_vec->length,
1031 (unsigned long)page_vec->pages[0],
1032 (unsigned long)page_vec->data_size,
1033 (unsigned int)mem_reg->lkey, mem_reg->mem_h,
1034 (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
1039 * Unregister (previosuly registered using FMR) memory.
1040 * If memory is non-FMR does nothing.
1042 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
1043 enum iser_data_dir cmd_dir)
1045 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1051 iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
1053 ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
1055 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
1060 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
1061 enum iser_data_dir cmd_dir)
1063 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1064 struct iser_conn *iser_conn = iser_task->iser_conn;
1065 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1066 struct fast_reg_descriptor *desc = reg->mem_h;
1072 spin_lock_bh(&ib_conn->lock);
1073 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
1074 spin_unlock_bh(&ib_conn->lock);
1077 int iser_post_recvl(struct iser_conn *iser_conn)
1079 struct ib_recv_wr rx_wr, *rx_wr_failed;
1080 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1084 sge.addr = iser_conn->login_resp_dma;
1085 sge.length = ISER_RX_LOGIN_SIZE;
1086 sge.lkey = ib_conn->device->mr->lkey;
1088 rx_wr.wr_id = (uintptr_t)iser_conn->login_resp_buf;
1089 rx_wr.sg_list = &sge;
1093 ib_conn->post_recv_buf_count++;
1094 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1096 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1097 ib_conn->post_recv_buf_count--;
1102 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1104 struct ib_recv_wr *rx_wr, *rx_wr_failed;
1106 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1107 unsigned int my_rx_head = iser_conn->rx_desc_head;
1108 struct iser_rx_desc *rx_desc;
1110 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1111 rx_desc = &iser_conn->rx_descs[my_rx_head];
1112 rx_wr->wr_id = (uintptr_t)rx_desc;
1113 rx_wr->sg_list = &rx_desc->rx_sg;
1115 rx_wr->next = rx_wr + 1;
1116 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1120 rx_wr->next = NULL; /* mark end of work requests list */
1122 ib_conn->post_recv_buf_count += count;
1123 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1125 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1126 ib_conn->post_recv_buf_count -= count;
1128 iser_conn->rx_desc_head = my_rx_head;
1134 * iser_start_send - Initiate a Send DTO operation
1136 * returns 0 on success, -1 on failure
1138 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1142 struct ib_send_wr send_wr, *send_wr_failed;
1144 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1145 tx_desc->dma_addr, ISER_HEADERS_LEN,
1148 send_wr.next = NULL;
1149 send_wr.wr_id = (uintptr_t)tx_desc;
1150 send_wr.sg_list = tx_desc->tx_sg;
1151 send_wr.num_sge = tx_desc->num_sge;
1152 send_wr.opcode = IB_WR_SEND;
1153 send_wr.send_flags = signal ? IB_SEND_SIGNALED : 0;
1155 ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1157 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1163 * is_iser_tx_desc - Indicate if the completion wr_id
1164 * is a TX descriptor or not.
1165 * @iser_conn: iser connection
1166 * @wr_id: completion WR identifier
1168 * Since we cannot rely on wc opcode in FLUSH errors
1169 * we must work around it by checking if the wr_id address
1170 * falls in the iser connection rx_descs buffer. If so
1171 * it is an RX descriptor, otherwize it is a TX.
1174 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1176 void *start = iser_conn->rx_descs;
1177 int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1179 if (wr_id >= start && wr_id < start + len)
1186 * iser_handle_comp_error() - Handle error completion
1187 * @ib_conn: connection RDMA resources
1188 * @wc: work completion
1190 * Notes: We may handle a FLUSH error completion and in this case
1191 * we only cleanup in case TX type was DATAOUT. For non-FLUSH
1192 * error completion we should also notify iscsi layer that
1193 * connection is failed (in case we passed bind stage).
1196 iser_handle_comp_error(struct ib_conn *ib_conn,
1199 void *wr_id = (void *)(uintptr_t)wc->wr_id;
1200 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
1203 if (wc->status != IB_WC_WR_FLUSH_ERR)
1204 if (iser_conn->iscsi_conn)
1205 iscsi_conn_failure(iser_conn->iscsi_conn,
1206 ISCSI_ERR_CONN_FAILED);
1208 if (is_iser_tx_desc(iser_conn, wr_id)) {
1209 struct iser_tx_desc *desc = wr_id;
1211 if (desc->type == ISCSI_TX_DATAOUT)
1212 kmem_cache_free(ig.desc_cache, desc);
1214 ib_conn->post_recv_buf_count--;
1219 * iser_handle_wc - handle a single work completion
1220 * @wc: work completion
1222 * Soft-IRQ context, work completion can be either
1223 * SEND or RECV, and can turn out successful or
1224 * with error (or flush error).
1226 static void iser_handle_wc(struct ib_wc *wc)
1228 struct ib_conn *ib_conn;
1229 struct iser_tx_desc *tx_desc;
1230 struct iser_rx_desc *rx_desc;
1232 ib_conn = wc->qp->qp_context;
1233 if (likely(wc->status == IB_WC_SUCCESS)) {
1234 if (wc->opcode == IB_WC_RECV) {
1235 rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
1236 iser_rcv_completion(rx_desc, wc->byte_len,
1239 if (wc->opcode == IB_WC_SEND) {
1240 tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
1241 iser_snd_completion(tx_desc, ib_conn);
1243 iser_err("Unknown wc opcode %d\n", wc->opcode);
1246 if (wc->status != IB_WC_WR_FLUSH_ERR)
1247 iser_err("wr id %llx status %d vend_err %x\n",
1248 wc->wr_id, wc->status, wc->vendor_err);
1250 iser_dbg("flush error: wr id %llx\n", wc->wr_id);
1252 if (wc->wr_id != ISER_FASTREG_LI_WRID &&
1253 wc->wr_id != ISER_BEACON_WRID)
1254 iser_handle_comp_error(ib_conn, wc);
1256 /* complete in case all flush errors were consumed */
1257 if (wc->wr_id == ISER_BEACON_WRID)
1258 complete(&ib_conn->flush_comp);
1263 * iser_cq_tasklet_fn - iSER completion polling loop
1264 * @data: iSER completion context
1266 * Soft-IRQ context, polling connection CQ until
1267 * either CQ was empty or we exausted polling budget
1269 static void iser_cq_tasklet_fn(unsigned long data)
1271 struct iser_comp *comp = (struct iser_comp *)data;
1272 struct ib_cq *cq = comp->cq;
1273 struct ib_wc *const wcs = comp->wcs;
1274 int i, n, completed = 0;
1276 while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
1277 for (i = 0; i < n; i++)
1278 iser_handle_wc(&wcs[i]);
1281 if (completed >= iser_cq_poll_limit)
1286 * It is assumed here that arming CQ only once its empty
1287 * would not cause interrupts to be missed.
1289 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1291 iser_dbg("got %d completions\n", completed);
1294 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1296 struct iser_comp *comp = cq_context;
1298 tasklet_schedule(&comp->tasklet);
1301 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1302 enum iser_data_dir cmd_dir, sector_t *sector)
1304 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1305 struct fast_reg_descriptor *desc = reg->mem_h;
1306 unsigned long sector_size = iser_task->sc->device->sector_size;
1307 struct ib_mr_status mr_status;
1310 if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1311 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1312 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1313 IB_MR_CHECK_SIG_STATUS, &mr_status);
1315 pr_err("ib_check_mr_status failed, ret %d\n", ret);
1319 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1320 sector_t sector_off = mr_status.sig_err.sig_err_offset;
1322 do_div(sector_off, sector_size + 8);
1323 *sector = scsi_get_lba(iser_task->sc) + sector_off;
1325 pr_err("PI error found type %d at sector %llx "
1326 "expected %x vs actual %x\n",
1327 mr_status.sig_err.err_type,
1328 (unsigned long long)*sector,
1329 mr_status.sig_err.expected,
1330 mr_status.sig_err.actual);
1332 switch (mr_status.sig_err.err_type) {
1333 case IB_SIG_BAD_GUARD:
1335 case IB_SIG_BAD_REFTAG:
1337 case IB_SIG_BAD_APPTAG:
1345 /* Not alot we can do here, return ambiguous guard error */